A Stability- and Aggregation-Based Method for Heart Rate Estimation Using Photoplethysmographic Signals During Physical Activity

TL;DR

This paper introduces a simple, effective method for estimating heart rate from PPG signals during physical activity, outperforming deep learning models without requiring complex computations or training data.

Contribution

A novel, one-size-fits-all signal processing method for heart rate estimation that avoids deep learning and reduces computational needs.

Findings

The proposed method outperforms CNN ensemble models on the PPG-DaLiA and IEEE_Test datasets.

It reduces mean absolute error by 1.45 bpm and 5.71 bpm compared to deep learning approaches.

The method works across multiple datasets and devices, including a new smart ring dataset.

Abstract

In recent years, the use of photoplethysmography (PPG)-based heart rate detection has gained considerable attention as a cost-effective alternative to conventional electrocardiography (ECG) for applications in healthcare and fitness tracking. Although deep learning methods have shown promise in heart rate estimation and motion artifact removal from PPG signals recorded during physical activity, their computational requirements and need for extensive training data make them less practical for real-world conditions when ground truth data is unavailable for calibration. This study presents a one-size-fits-all approach for heart rate estimation during physical activity that employs aggregation-based techniques to track heart rate and minimize the effects of motion artifacts, without relying on complex machine learning or deep learning techniques. We evaluate our method on four publicly available datasets—PPG-DaLiA, WESAD, IEEE_Training, and IEEE_Test, all recorded using wrist-worn devices—along with a new dataset, UTOKYO, which includes PPG and accelerometer data collected from a smart ring. The proposed method outperforms the CNN ensemble model for the PPG-DaLiA dataset and the IEEE_Test dataset and reduces the mean absolute error (MAE) by 1.45 bpm and 5.71 bpm, respectively, demonstrating that effective signal processing techniques can match the performance of more complex deep learning models without requiring extensive computational resources or dataset-specific tuning.

Figures